CARE helps VLA robots recover from failures, boosting task success by up to 15.9 points

dalian-university-of-technology · hf · 2026-09-22

Researchers from Dalian University of Technology propose CARE (Corrective Atomic Robotic Execution), tackling the brittleness of Vision-Language-Action policies when execution deviates from nominal trajectories.

Instead of generating corrective data from manual or random perturbations, CARE collects failed rollouts, models stage-conditioned post-failure deviations, and synthesizes representative failure states and corrective demonstrations. At inference, it combines stage-wise planning with physically grounded 3D monitoring to trigger atomic adjustments while preserving task progress.

They also release the Failure State Recovery Benchmark (FSR-Bench). Across multiple VLA backbones, simulation benchmarks, and real dual-arm tasks, CARE delivers average task-success gains of 14.5 points in simulation and 15.9 points in the real world. Code, models, and data are open-sourced.

Original post →

More from Embodied

Embodied channel →